LHCbPR - Performance and Regression Tests


LHCbPR is responsible for systematically running the regression tests, collecting and comparing results of these tests so that any changes between different setups can be easily observed. The framework is based on a microservices architecture which breaks a project into loosely coupled modules communicating with each other through APIs. Most of the developed modules are generic which means that the proposed framework can be adopted for other experiments.


The sequence diagram of the LHCbPR is shown in this figure: https://twiki.cern.ch/twiki/pub/LHCb/LHCbPR/SequenceDiagram.pdf.


Configuration of the test

The test is configured via https://gitlab.cern.ch/lhcb-core/LHCbNightlyConf/blob/master/test_schedule2.xml. The example of the configuration is the following:

    <schedule type="week" time="10:00">Mon,Tue,Wed,Thu,Fri</schedule>
    <test runner="lhcbpr" group="MiniBrunel" env="lb-run-gaudirun|TimeLineHandler"/>

The meaning of the keys:

  • schedule type : you can specify week or month
  • schedule time : usually ignored, see LHCbPR#Launching_the_tests below.
  • schedule : the day of the week if type="week" (always respected).
  • project : name of the project to be tested
  • slot, platform : you can use regex, e.g. lhcb-master|lhcb-head, also globbing for platform, e.g. x86_64-slc6-gcc*-opt
  • test runner : use lhcbpr
  • test group : description of the option file, please contact us to add it to LHCbPR database https://lblhcbpr.cern.ch/api/options/
  • test env : description of the command to run the test and the name of the handler. Please contact us to add the executable to LHCbPR database https://lblhcbpr.cern.ch/api/executables in case it is not already there. After | please specify the list of handlers separated by comma. The name should correspond to the file committed to https://gitlab.cern.ch/lhcb-core/LHCbPR2HD
  • os_label : use perf for slc6 tests or perf-centos7 for centos7 or perf-centos7-timing for the timing tests on centos7 (the dedicated machine with the single executor in jenkins will be used)
  • count : specify number of runs for the test (you can run multiple tests to check statistics - standard deviation will be computed for the extracted metrics)


Here you can find the dashboard for the periodic tests. One can verify here the status of the executed tests. The colour code is the following:

  • running tests in blue
  • successful tests in green
  • failed tests in red
  • tests which have been executed with success, but the handler failed (so that there is no output results for LHCbPR) in yellow
The urls to log files of the test (stored on EOS) and output of the jenkins job are provided.

Launching the tests

  • Automatic starting. Tests defined in configuration file are started automatically by the Jenkins job. The job checks (messaging infrastructure described here is used for this purpose) every 5 minutes if there any new builds for the defined tests for the given day (the exact time from the configuration is ignored). If yes, another Jenkins job is triggered (the same which is used for tests of nightly builds) which actually runs the test. If the messaging infrastructure is down, as a backup, the tests may be started according to the time in the configuration when tests-poll Jenkins jobs will be enabled.
  • Manual start on demand. After login to the dashboard, there is an orange button in top right, called Start new periodic test. After clicking it, you need to provide the same information which is in configuration file (except for scheduling). The message will be sent to the queue checked by Jenkins job every 5 minutes. Alternatively you can use command line:
    export RMQPWD=lhcbpr/lhcbpr
    lbq-requesttest <slot> <buildid> <project> <config> <group> <env>
    For example:
    lbq-requesttest 1467 lhcb-sim09  Gauss x86_64-slc6-gcc49-opt "GAUSS-RADLENGTHSCAN" "lb-run|RadLengthHandler"
    One can use also -l flag to specify the machine label in jenkins.


Here are the machines used for running the periodic tests:

  • volhcb05 with SLC6 (8 executors)
  • lblhcbpr6 with Centos7 (8 executors)
  • lblhcbpr1 with Centos7 (1 executor)
  • lbhltperf01 node devoted for throughput testing


Handlers are python scripts used to extract relevant information from the the output produced by the test runs. The BaseHandler class enables to save Int, Float, String, JSON and File. The LHCbPR framework produces the zip file with the collected results which is sent to database through Dirac Storage element (/lhcb/prdata/zips). The description how to create handler and test it can be found here.


The results of the tests are automatically picked up by web front-end available here. For the development of the specific analysis module please see here. The generic tool to compare the plots can be found by going to LHCbPR Jobs and ROOT file viewer tabs. To perform the trend analysis see the example below.

There is an ongoing work on the new version of the web front-end available here.

Results of the throughput tests are available here.

Example of plotting trend

Letís assume we want to plot the time spent by EVENT_LOOP as a function of the software version. The option file can be found here, and the command used to run the test is:

lb-run --use=PRConfig -c x86_64-slc6-gcc62-opt --user-area=$(pwd)/../build Brunel/HEAD gaudirun.py \$PRCONFIGOPTS/Brunel/PRTEST-COLLISION15-1000evts.py 

To plot the trend:

  • go to Trends/Trends tab and select Brunel from the list of applications
  • select the option you are interested in, in this case PRTEST-COLLISION15-1000
  • tick Show Nightly versions and specify the number of versions to show, e.g. 50.
  • start typing the name of the algorithm in the field Filter attributes and click Show
You should see the plot: https://twiki.cern.ch/twiki/pub/LHCb/LHCbPR/BrunelTiming.pdf.

Requirements for participation

To add your project to LHCbPR the following information is needed:

  • Command to run the test
  • The option file stored e.g. in PRConfig
  • Handler to extract the relevant information from the test
  • Analysis module in case you are interested in specific presentation of results (other than trend analysis and generic comparison of plots using ROOT file browser)

Analysing data from LHCbPR in SWAN

Since May 2018, results of LHCbPR tests are copied to Hadoop Distributed File System (HDFS), see the user guide, twiki and knowledge base articles for reference. Details of the procedure can be found in https://gitlab.cern.ch/maszyman/lhcbpr-hadoop.

Owing to that, one can use SWAN notebooks to create custom reports on test results. Examples are available in /eos/user/m/maszyman/SWAN_projects/read_hdfs. The requirement to read from HDFS is to belong to ai-hadoop-users e-group (which can be granted by opening a SNOW ticket to the Hadoop and Spark Service to request access).

To be able to read from HDFS, go to SWAN, open a new notebook (you may need to create a new project first), click on a star (second to last in top row - Spark clusters connection). By default you should be directed to analytix cluster, when you click 'Connect'.

Alternatively, you can use docker container to run pyspark:

# get docker image
docker login gitlab-registry.cern.ch
docker pull gitlab-registry.cern.ch/db/cerndb-infra-hadoop-conf:qa

# run it
docker run -d -it -p 5000-5300:5000-5300 --hostname $HOSTNAME --name "lhcbpr-hadoop"  -v /cvmfs/sft.cern.ch:/cvmfs/sft.cern.ch:shared gitlab-registry.cern.ch/db/cerndb-infra-hadoop-conf:qa

# go into docker image
docker exec -it lhcbpr-hadoop bash

# get kerberos token
kinit ${USER}@CERN.CH

# run pyspark interactively (alternatively write a script and run it using spark-submit command)

The notebook producing trend plots of the throughput is available here.

Setup of the throughput tests

The throughput tests are running making use of lhcb-benchmark-scripts repository. Please see readme for details.

You can find an example of the setup in https://lblhcbpr.cern.ch/api/options/90/. Please note that lhcb-benchmark-scripts repository is setup in LbNightlyTools.


For the development of LHCbPR, please see here.

Contact and Infos




Web application:

PR dashboards:

The output of PR jobs can be shown on PR dashboards. Some links to these dashboards are: More information on the reconstruction and Allen dashboards is available here: https://gitlab.cern.ch/lhcb-rta/data-challenges.

API service:

ROOT HTTP service:

Testsí output handlers:

Project builder:

Jenkins configuration

Configuration of the periodic tests

Collection of various talks given on the subject of LHCbPR

Topic attachments
I Attachment History Action Size Date Who Comment
PDFpdf BrunelTiming.pdf r1 manage 821.8 K 2017-05-22 - 16:17 MaciejSzymanski  
PDFpdf SequenceDiagram.pdf r1 manage 79.1 K 2017-05-22 - 16:17 MaciejSzymanski  
Edit | Attach | Watch | Print version | History: r22 < r21 < r20 < r19 < r18 | Backlinks | Raw View | WYSIWYG | More topic actions
Topic revision: r22 - 2021-04-27 - MaciejSzymanski
    • Cern Search Icon Cern Search
    • TWiki Search Icon TWiki Search
    • Google Search Icon Google Search

    LHCb All webs login

This site is powered by the TWiki collaboration platform Powered by PerlCopyright &© 2008-2023 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
or Ideas, requests, problems regarding TWiki? use Discourse or Send feedback